Sentient Futures
Back to proposals

WelfareGuard: An open-source eval-to-guardrail pipeline for nonhuman welfare in agentic AI systems

commitment
10 hrs/week
flexibility
Open to proposals
mentored by
Vidura Wijekoon
Vidura WijekoonSenior AIML Engineer/MSc Researcher in AI and Data Science

The project

As AI agents increasingly make autonomous decisions in supply chains, agriculture, content generation, and research, harms to nonhuman sentient beings (animals today, potentially digital minds tomorrow) go largely unmeasured and unmitigated. Existing welfare benchmarks measure model attitudes, but nothing translates those measurements into runtime protections for deployed agentic systems.

WelfareGuard extends my open-source AI safety library, aisafepy (on PyPI: capability-based information-flow control, streaming cascaded guardrails, and an eval-to-guardrail compiler supporting PyRIT, Garak, and Inspect), into the nonhuman welfare domain. Over 10 weeks we will build:
(1) a welfare-harm eval suite for agentic pipelines - testing not just what models say about animals/digital minds, but what multi-step agents do (tool calls, plans, delegated sub-agent behavior); (2) a compiler pass that converts failed welfare evals directly into deployable runtime guardrails; and
(3) a public demo agent + technical report showing measured harm-rate reduction.

Expected outputs: a released open-source package (v1 on PyPI), an eval dataset/harness compatible with Inspect, a technical report (EA Forum/LessWrong post + arXiv preprint), and a draft funding proposal (e.g., toward Longview's digital minds RFP) to continue the work beyond the incubator.

My prior BlueDot research on cross-agent deception detection (SAEs, linear probes, attribution graphs on LangGraph pipelines) provides the interpretability backbone for a stretch goal: probing whether agents' internal representations diverge from their stated reasoning on welfare-relevant decisions.

The mentee's role

Three complementary workstreams, one per mentee (with pairing on integration weeks):

  1. Evals engineer - design and implement the agentic welfare-harm eval scenarios (prompt suites, tool-use traps, multi-agent setups) in Inspect; run baseline measurements across 3⁠–⁠4 frontier models.
  2. Guardrails engineer - extend the aisafepy eval-to-guardrail compiler to the new eval formats; build and benchmark the runtime mitigation layer (latency, false-positive rate, harm-rate reduction).
  3. Research & comms lead - literature review connecting animal welfare / digital minds frameworks to concrete agent failure modes; co-author the technical report; draft the continuation funding proposal.

Sample weekly tasks:

  • Writing eval scenarios
  • Running experiments
  • PR reviews on the open-source repo
  • Drafting report sections

All work happens in a public GitHub repo, so mentees leave with citable, visible contributions.

Who I'm looking for

Must-haves:

  • Intermediate Python and comfort with Git/GitHub workflows
  • Genuine interest in nonhuman welfare (animals or digital minds)
  • Ability to work independently between weekly syncs

Nice-to-haves:

  • Experience with LLM APIs or agent frameworks (LangGraph/LangChain)
  • Prior eval-writing (Inspect, PyRIT, Garak)
  • Interpretability exposure (SAEs, probes)
  • Technical writing experience

Roles are differentiated, so a strong writer/researcher with lighter coding skills fits workstream 3.

Questions for applicants

  1. Which of the three workstreams (evals, guardrails, research & comms) best fits your skills, and why?
  2. Link to a code repository, technical writing sample, or project you're proud of.
  3. Describe a time you shipped something end-to-end with minimal supervision.

Support offered

  • Hands-on engineering mentorship (architecture reviews, PR feedback, debugging sessions) from 7+ years of production ML across 52+ shipped AI projects
  • Domain expertise in AI safety tooling, agentic security (co-author of OWASP Securing Agentic Applications Guide sections on memory integrity and trust boundaries), and mechanistic interpretability
  • Project scoping and accountability structures honed from managing an 10-engineer team and mentoring 500+ people across 30+ countries
  • Connections into AI safety and security communities (OWASP GenAI, BlueDot network, AI SGE Chapter Sri Lanka) for distribution and feedback on outputs
format
Product & entrepreneurship
topic
Post-AGI transition · Tools for advocates · Artificial sentience
Vidura Wijekoon

Vidura Wijekoon

Senior AIML Engineer/MSc Researcher in AI and Data Science

Most enterprise AI projects die in the compliance review, not the Jupyter notebook. I build the ones that survive both.

I'm Vidura, a Senior AI/ML Engineer & Line Manager at Virtusa, specializing in production GenAI systems where safety, governance, and ethics aren't afterthoughts- they're architecture decisions. If your AI passed demo day but is stuck in legal review, procurement limbo, or a "who owns this model?" debate, we speak the same language.

Over seven years, I've shipped 50+ AI/ML projects, led solution architecture for GenAI and ML applications on AWS and GCP, and managed engineering teams through the messy reality of taking models from prototype to production in regulated environments.

Before Virtusa, I co-founded Trinet Innovations as COO, where I learned that the hardest AI problem isn't the model, it's translating technical capability into business value that boards and compliance teams will actually approve. That lesson shapes everything I build today.

A few proof points that matter more than certifications:

  • Contributed to consciousness-AI research on Hanson Robotics' Sophia through Omdena's global research team
  • U.S. Department of State Professional Fellow, selected to represent AI innovation from Sri Lanka
  • IEEE PES Award recipient for power-sector AI research
  • Operational Lead, AI SGE Chapter Sri Lanka, driving regional AI knowledge-sharing and community growth
  • Delivered 62+ technical workshops and mentored 500+ engineers, students, and startup founders across 30+ countries
  • GCP AWS Certified ML Specialist | Certified AI Security & Governance Practitioner
  • MSc researcher in AI & Data Science, University of Sri Jayewardenepura

Three reasons people land on this profile:

  1. You're an enterprise leader wondering if your GenAI initiative will survive procurement, legal, and audit. DM me, and I'll offer a 3⁠0⁠-⁠m⁠i⁠n⁠u⁠t⁠e scoping call.
  2. You're a founder building an AI product and need a fractional AI lead or architecture review. My calendar link is in Featured.
  3. You're an engineer or student who wants to build AI that actually ships. My newsletter and free playbooks are pinned at the top of my profile; start there.

I come from an AI + Power Engineering background, with astrophysics and cosmology as a hobby field. Toggling between grid stability challenges about how galaxies hold themselves together built an early habit of respecting the physics before trusting the model. That cross-domain instinct now shapes how I apply AI across power systems, healthcare, and whatever regulated problem lands on my desk next.